Detection of Rail Fasteners from Aerial Images Using Deep Convolution Neural Networks

Reliable rail infrastructure helps in ensuring a state's economic stability. Timely detection of missing rail fasteners is crucial in avoiding fatal accidents and plays a significant role in rail safety assurance. In this paper we propose the use of one-stage RetinaNet deep convolution network architecture to detect rail fasteners on rail tracks by analyzing the aerial imagery. Quantitative assessment was conducted in a one km railway stretch by acquiring aerial images of size 5472×3078, using a DJI Phantom 4 from a height of 100m. However, the background complexity renders it as a demanding task. Results demonstrate that the learned deep features promise a robust method, if trained on a larger dataset and can help identify rail anomalies which may require imminent repair.